Research map / Video / Video generation
Learning Long-term Motion Embeddings for Efficient Kinematics Generation
Nick Stracke, Kolja Bauer, Stefan Andreas Baumann, Miguel Ángel Bautista, Josh Susskind, Björn Ommer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| raise temporal compression of the motion embedding (2x to 64x) under fixed compute (ablation) | ZipMo · Pexels motion generation | motion generation quality and throughput | Improved |
| raise temporal compression factor (ablation) | ZipMo · SSv2 subset | kNN accuracy | Improved |
| generate in learned motion latent space instead of video generation (Wan) | ZipMo · Pexels (sample matched) | EPE: 4.68 -> 1.17 | Improved |
| generate in motion latent space instead of Wan under matched wall-clock time | ZipMo · Pexels (time matched) | Min MSE: 64.20 -> 21.29 | Improved |
| generate in motion latent space instead of Motion-I2V flow | ZipMo · Poked motion, Dense | EPE: 8.8 -> 1.1 | Improved |
| add track dropout to training tracks (ablation) | ZipMo · Reconstruction accuracy | delta avg: 96.8 -> 94.0 | Hurt |
| train only on non-occluded tracks (filter occlusions) (ablation) | ZipMo · Reconstruction accuracy | delta avg: 96.8 -> 93.2 | Hurt |
| train on CoTracker3 instead of TapNext, evaluate on TapNext (ablation) | ZipMo · Cross-tracker reconstruction | delta avg: 96.8 -> 96.3 | No clear effect |
| arrange latent tokens in a grid with 2D RoPE (ablation) | ZipMo · Reconstruction PCK | PCK | Improved |
| track generated Wan video with TapNext instead of CoTracker3 | Wan · Pexels | Min MSE: 29 -> 37 | Hurt |
| use ZipMo planner instead of ATM | ZipMo · LIBERO Goal | success rate: 77.8 -> 69.7 | Hurt |
| use ZipMo planner instead of Amplify | ZipMo · LIBERO 90 | success rate: 66.0 -> 62.3 | Hurt |
Closest papers by meaning
- HIMMEL: Hierarchical Interleaved Multi-stream Motion Encoding for Long Video Understanding (NeurIPS 2026 · similarity 0.88)
- FlowMo: Variance-Based Flow Guidance for Coherent Motion in Video Generation (NeurIPS 2025 · similarity 0.87)
- VideoJAM: Joint Appearance-Motion Representations for Enhanced Motion Generation in Video Models (ICML 2025 · similarity 0.87)
- MotionRAG: Motion Retrieval-Augmented Image-to-Video Generation (NeurIPS 2025 · similarity 0.87)
- DynaTokens: Teaching Dynamics to Camera-Controlled Video Models at Test Time (NeurIPS 2026 · similarity 0.87)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
